Identifying Sociodemographic, Clinical, and Behavioral Determinants of Self- Reported Kidney Stone History in US Adults: A Population-Based Comparative Analysis of Logistic Regression, Random Forest, and XGBoost Using NHANES 2007–2018
Abstract Background Logistic regression remains conventional for identifying correlates of self-reported kidney stone history, but multidimensional comparisons with machine learning—integrating discrimination, calibration, and clinical utility—remain scarce. This study aimed to identify determinants of self-reported kidney stone history among US adults and to compare logistic regression, random forest, and XGBoost. Methods We conducted a pooled cross-sectional analysis of six consecutive NHANES cycles (2007–2018), including adults aged ≥ 20 years with complete data on kidney stone history and 16 prespecified candidate predictors. LASSO logistic regression with 10-fold cross-validation selected variables in a 70% training subset. Logistic regression, random forest, and XGBoost were trained on the identical predictor set and evaluated on the held-out 30% test set by discrimination (AUC with DeLong's test), calibration (intercept and slope), and clinical utility (decision curve analysis). Survey-weighted sensitivity analysis, collinearity assessment, nomogram, and SHAP analyses. Results Of 25,560 adults, 2,462 (9.6%; survey-weighted 9.99%) reported a lifetime history of kidney stones. LASSO retained 14 of 16 candidates. Eight variables were independently associated with kidney stone history: older age, male sex, lower BMI (inverse), higher waist circumference, elevated blood urea nitrogen, hypertension, diabetes mellitus, and current smoking (all 95% CIs excluding unity). BMI and waist circumference showed substantial collinearity (VIFs 7.69 and 7.81). Logistic regression achieved an AUC of 0.644, comparable to XGBoost (0.641; DeLong P = 0.591) and superior to random forest (0.617; P = 0.0005), with the strongest calibration (intercept − 0.147; slope 0.892), followed by XGBoost (slope 0.771) and random forest (slope 0.569). On decision curve analysis, random forest's net benefit fell below the treat-none line within the 15–35% threshold range, whereas logistic regression and XGBoost remained stable. SHAP analysis confirmed age, waist circumference, and blood urea nitrogen as the top contributors; survey-weighted results were consistent. Conclusions Older age, male sex, adiposity-related traits, elevated blood urea nitrogen, hypertension, diabetes mellitus, and current smoking were independently associated with prevalent kidney stone history. Logistic regression matched XGBoost in discrimination while showing the best calibration, stable clinical net benefit, and full interpretability. Given the cross-sectional, self-reported design, these findings reflect correlates rather than causal predictors and warrant external validation.
Authors
- Yanzhu Wang (ORCID: https://orcid.org/0000-0002-5656-7597)
- Li Chengbin
- Dawei Zhao¹
- Hao Wu¹
Publication Details
- Journal
- Research Square
- Published
- 2026-11-16
- DOI
- https://doi.org/10.21203/rs.3.rs-11445749/v1
- Primary Topic
- Kidney Stones and Urolithiasis Treatments
- Type
- preprint